Inverse relationship between image matching and feed-forward reconstruction

Investigate what image matchers know about feed-forward 3D reconstruction, thereby characterizing the inverse relationship to the use of feed-forward 3D models for image matching.

Background

The paper studies what feed-forward 3D reconstruction models, particularly VGGT-$\Omega$, encode about image matching and demonstrates that their learned representations and raw geometric predictions can support matching. The authors identify the inverse direction—examining whether and how image matchers encode information useful for feed-forward reconstruction—as a separate question that is outside the scope of the reported experiments and remains a possible avenue for future investigation.